TOON: “JSON for AI” (is it any good?)

Next episode : GOON

I love XML prompting. It burns so much investors money.

This feels like a "problem" that should be "solved" at the tokeniser level.

CSV looks like it uses the least amount of tokens

Add tip of the day to t3 chat for such insights: do not use claude with large context - use gemini 2.5 flash

Wait till they discover vsc standard : values seperated by commas it will revolutionize the process of reducing the token count.

I used TOON for a take home interview project a couple of weeks ago. I had the exact use case of passing tabular JSON data into a prompt to generate another sample JSON data output. When I tested with JSON, I'd get an error in the JSON output ~10% of the time. When I switched to TOON and didn't see any errors in my output at all. This is just anecdotal, but it's certainly useful in my experience. (And I got the job!)

Super Prompt Oriented Object Notation sounds nice.

My favorite thing to do nowadays is change the prompt builder function in any ai powered app to just print the text out like a normal list instead of JSON and then watch as it does it just as well for 25% less cost and processing time.

I like how it looks! It is visually pleasant to see less repeating characters of JSON :D